USPEX-ML

Official Resources

• Homepage: https://uspex-team.org/en/uspex/
• Source Repository: https://uspex-team.org (distributed upon request)
• Documentation: https://uspex-team.org/en/uspex/documentation
• License: Free for academic use (registration required)

Overview

USPEX-ML refers to the machine learning-enhanced functionality of USPEX (Universal Structure Predictor: Evolutionary Xtallography), the widely used evolutionary algorithm for crystal structure prediction. USPEX, developed by A.R. Oganov and collaborators since 2004, employs evolutionary algorithms to explore the energy landscape and predict stable crystal structures. Recent versions integrate machine learning interatomic potentials (MLIPs) to dramatically accelerate structure relaxation and energy evaluation during the evolutionary search.

USPEX 25, the latest release, includes built-in deep learning tools (MatterSim) for fast crystal structure relaxation, enabling users to start projects immediately without external quantum-mechanical software. ML potentials such as MTP (Moment Tensor Potential), GAP (Gaussian Approximation Potential), DeepMD, and Allegro can be integrated as external energy evaluators, replacing expensive DFT calculations during the initial generations of the evolutionary search. This ML acceleration enables crystal structure prediction for larger and more complex systems, including molecular crystals with dozens or hundreds of atoms.

Scientific domain: Crystal structure prediction, materials discovery, machine learning
Target user community: Computational materials scientists predicting new crystal structures

Theoretical Methods

• Evolutionary algorithm (genetic algorithm) for global optimization
• Machine learning interatomic potentials (MLIPs) for energy evaluation
• MatterSim built-in deep learning model for structure relaxation
• Multi-stage relaxation with ML pre-relaxation and DFT refinement
• Variable-composition and fixed-composition searches
• Evolutionary operators: heredity, mutation, permutation, lattice mutation
• Topology-based structure generation
• Finite-temperature crystal structure prediction with ML potentials

Capabilities (CRITICAL)

• Crystal structure prediction with ML acceleration
• Built-in MatterSim deep learning for fast relaxation
• Integration with external ML potentials (MTP, GAP, DeepMD, Allegro)
• Variable-composition and fixed-composition searches
• Molecular crystal prediction
• Finite-temperature structure prediction with anharmonic free energy
• Multi-stage relaxation workflow (ML → DFT)
• No MATLAB required in USPEX 25
• Windows and Linux support
• HPC integration for remote cluster submission
• Visualization tool STMng

Inputs & Outputs

Input formats:

• USPEX input files (Python-based in v25)
• Chemical composition and stoichiometry ranges
• ML potential or DFT code selection
• Evolutionary algorithm parameters (population, generations, operators)

Output data types:

• Predicted stable and metastable crystal structures
• Convex hull diagrams
• Energy-composition phase diagrams
• Structure-property predictions
• Relaxation trajectories

Interfaces & Ecosystem

• Programming language: Python (v25), MATLAB (legacy)
• ML potentials: MatterSim (built-in), MTP, GAP, DeepMD, Allegro, ReaxFF-nn
• DFT backends: VASP, Quantum ESPRESSO, GULP, ORCA, DFTB+
• Platforms: Windows, Linux (no MATLAB required in v25)
• HPC: Remote cluster job submission

Limitations & Known Constraints

• Registration required for download
• ML potentials require training data for target chemical systems
• Quality of predictions depends on ML potential accuracy
• Evolutionary algorithm parameters need tuning for complex systems
• Very large unit cells still challenging even with ML acceleration

Performance Characteristics

• ML potentials provide 10^3-10^6 speedup over DFT for energy evaluation
• Multi-stage relaxation: ML pre-relaxation reduces DFT calculations needed
• Population size and generations can be increased with ML acceleration
• Parallel evaluation of structures in each generation
• USPEX 25 runs on PC without MATLAB runtime

Comparison with Other Codes

• vs COPEX: COPEX runs co-evolutionary USPEX processes; USPEX-ML uses ML for single-process acceleration
• vs CALYPSO: CALYPSO uses particle swarm optimization; USPEX uses evolutionary algorithm with ML
• vs AIRSS: AIRSS uses random search; USPEX-ML uses informed evolutionary search with ML
• vs Crystal Diffusion models: Diffusion models generate structures directly; USPEX-ML optimizes via evolution

Best Practices

• Use ML potentials for initial generations, switch to DFT for final refinement
• Train ML potentials on representative DFT data for target system
• Use multi-stage relaxation for best efficiency
• Run sufficient generations for convergence
• Analyze convex hull for thermodynamic stability
• Use built-in MatterSim for quick starts without DFT setup

Verification & Sources

Primary sources:

1. USPEX website: https://uspex-team.org/en/uspex/
2. USPEX release notes: https://uspex-team.org/en/uspex/release
3. A.R. Oganov and C.W. Glass, J. Chem. Phys. 124, 244704 (2006)
4. USPEX-ML applications: J. Mater. Inform. (2024) - CL-20/HMX cocrystals

Confidence: VERIFIED - Official website and release notes confirmed accessible